Cesar Hidalgo

Foreword

By Edo Segal

The question that haunted me after Trivandrum was not about productivity.

Productivity I could measure. Twenty-fold. Real. Verified across sprints and shipping dates and features that worked. The dashboards confirmed everything I wanted to believe. My team was faster, bolder, reaching into domains they had never touched. The numbers were extraordinary.

But numbers measure what you produce. They do not measure what you keep.

I started noticing something I could not quantify. Engineers who built features with Claude in hours could not explain those features without Claude the next morning. Not because they were careless. Because the knowledge had passed through them the way water passes through a pipe. The pipe is not wet afterward. The output was real. The understanding was borrowed.

I did not have a word for this until I encountered César Hidalgo's work. Hidalgo is a physicist who became an economist by asking a question that economists had been avoiding: Why do some countries stay poor despite having access to the same information as rich ones? His answer was not about money or resources or even education in the conventional sense. It was about something he called crystallization — the process by which knowledge escapes individual minds and embeds itself in objects, institutions, and systems that persist independently of any single knower.

A hammer crystallizes metallurgy. A compiler crystallizes computation. A large language model crystallizes the connective tissue between every domain of human thought ever committed to text.

That last crystallization is what we are living through. And Hidalgo's framework reveals the thing the productivity metrics hide: the difference between knowledge you can access and knowledge you actually own. Between output that depends on a subscription and capability that survives when the subscription lapses.

This distinction matters because the entire premise of The Orange Pill — that AI is an amplifier, and the quality of the output depends on the quality of what you feed it — assumes there is something durable being fed. If the human side of the partnership is itself borrowed, if the judgment directing the tool was never sedimented through friction and failure and time, then what exactly is being amplified?

Hidalgo gave me a framework for measuring the thing I was afraid to look at directly. Not whether we were building faster. Whether what we were building would last.

That is why this book exists. Another lens. Another floor of the tower. Another crack in the fishbowl.

— Edo Segal ^ Opus 4.6

About Cesar Hidalgo

1979-present

César A. Hidalgo (1979–present) is a Chilean-Spanish-American physicist, information theorist, and author whose work has reshaped how economists and policymakers understand development, complexity, and the role of knowledge in national prosperity. Born in Santiago, Chile, he studied physics at the Pontificia Universidad Católica de Chile before completing his PhD at the University of Notre Dame. He held a faculty position at the MIT Media Lab for over a decade, where he led the Macro Connections group and co-created the Atlas of Economic Complexity, a pioneering tool mapping the productive knowledge of nations. His major books include *Why Information Grows: The Evolution of Order, from Atoms to Economies* (2015) and *How Humans Judge Machines* (2021). Hidalgo developed key concepts including personbytes (the finite amount of productive knowledge an individual can hold), knowledge crystallization (the embedding of know-how in objects and institutions), and economic complexity as a predictor of national growth. He has held positions at the University of Toulouse, Corvinus University of Budapest, and founded the Center for Collective Learning. In 2026, he launched JAIGP (Journal for AI Generated Papers) in collaboration with Claude, applying his framework to the institutional challenges of AI-era knowledge production. His work bridges physics, economics, network science, and information theory to argue that the wealth of nations is determined not by what they extract but by what they know how to make — and how durably that knowledge is embedded in their institutional fabric.

Chapter 1: The Crystallization of Know-How

A hammer is not a simple thing.

It looks simple. A handle, a head, a striking surface. Pick it up, swing it, drive a nail. A child can use one. But the hammer you hold in your hand is the endpoint of a crystallization process that spans millennia. Someone had to discover that certain rocks, heated past a specific threshold, yielded metal that could be shaped. Someone else had to learn that the ratio of carbon to iron determined whether the head would shatter on impact or absorb the blow and transfer its energy cleanly into the nail. Someone else had to work out the geometry of the grip — the angle, the length, the diameter that allows a human wrist to generate maximum force with minimum strain over thousands of repetitions. The hammer embodies all of this. The person swinging it embodies none of it.

This is what César Hidalgo means by crystallization: the process by which knowledge escapes the minds that produced it and takes up residence in objects, institutions, and systems that can be used by people who do not possess the knowledge themselves. In Why Information Grows, Hidalgo made the case that the entire history of economic development can be understood as a history of crystallization — the progressive embedding of human know-how into forms that persist independently of any individual knower. The economy is not, in this view, a system for allocating resources. It is a system for accumulating information. And the wealth of nations is determined not by what they extract from the ground but by what they have crystallized into the objects, organizations, and institutional arrangements that constitute their productive capacity.

A computer is a more elaborate crystallization. It embodies centuries of mathematical insight — Boolean algebra, information theory, the specific engineering decisions that determine how electrical charge is stored and manipulated in silicon. The user who opens a spreadsheet and calculates a sum is accessing an extraordinary density of crystallized knowledge. The crystallization is so complete, so seamlessly embedded in the artifact, that the user does not experience it as knowledge at all. It feels like a feature. It feels like the machine doing something. The knowledge has been compressed into an interface so thoroughly that the compression itself has become invisible.

A large language model represents a crystallization of a different order entirely.

It crystallizes not a specific domain of knowledge but the accumulated textual output of human civilization — the patterns of thought, the structures of argument, the relationships between concepts, the ways in which ideas have been connected and disconnected and reconnected across millions of documents and billions of words. When a user sits down with Claude and describes a problem in natural language, the user is accessing a crystallization so vast, so densely layered, that the metaphor of the hammer begins to strain. This is not a tool that embodies one domain of knowledge. It is a tool that embodies the connective tissue between domains.

In December 2025, a Google principal engineer sat down with Claude Code and described, in three paragraphs of plain English, a problem her team had spent a year trying to solve. One hour later, Claude had produced a working prototype. "I am not joking," she wrote publicly, "and this isn't funny."

Hidalgo's framework illuminates what actually happened in that hour. The engineer was not witnessing the invention of new knowledge. She was witnessing the decrystallization and recrystallization of existing productive knowledge. The knowledge required to build her team's system already existed — distributed across thousands of papers, codebases, architectural patterns, and engineering decisions that had been made by thousands of people over decades. What the language model had done was gather that distributed knowledge into a conversational interface. The engineer accessed the crystallization through natural language. The prototype emerged not from creation but from compression — the gathering of dispersed know-how into a single, navigable structure.

This compression is an extraordinary event in the history of human tool use. But Hidalgo's framework insists on a distinction that the exhilaration of the moment tends to obscure. Crystallization and understanding are different phenomena. They are related. They often co-occur. But they are not the same thing. The person who swings the hammer accesses crystallized metallurgical knowledge without understanding metallurgy. The person who uses the spreadsheet accesses crystallized mathematical knowledge without understanding information theory. In both cases, the output is real and the understanding is borrowed. The borrowing works because the tool is doing the knowing.

The Orange Pill describes an engineer in Trivandrum, India — a woman who had spent eight years working exclusively on backend systems and had never written a line of frontend code — who built a complete user-facing feature in two days using Claude. Not a prototype. A deployable feature with interface logic that responded to user interaction. The output was indistinguishable from what a frontend specialist would produce. But the knowledge structure behind the output was entirely different. The specialist's output would have rested on years of accumulated understanding — mental models of how the DOM works, intuitions about responsive layouts, pattern recognition built through hundreds of failed implementations. The Trivandrum engineer's output rested on crystallized knowledge accessed through conversation. She could not, after the experience, sit down without the tool and reproduce the work. The knowledge had not transferred to her. It had been accessed through her — the way water passes through a pipe without the pipe becoming water.

For certain kinds of work, this distinction does not matter. A person who needs a website does not need to understand the Document Object Model any more than a person who needs transportation needs to understand internal combustion. The crystallization of productive knowledge into tools that eliminate the need for understanding is, in fact, one of the great achievements of human civilization. Every tool performs this function. The hammer liberated the nail-driver from metallurgy. The compiler liberated the programmer from assembly language. Each liberation expanded the population of people who could produce useful output, and each expansion generated economic value that exceeded what the previous, smaller population of knowledgeable practitioners could generate.

But Hidalgo insists on a distinction between tools that crystallize knowledge for use and tools that crystallize knowledge for production. The hammer crystallizes metallurgical knowledge for the person who drives nails — a user. The compiler crystallizes computational knowledge for the person who writes programs — a producer. In each case, the tool operates at a level below the level at which the person is working. The nail-driver does not need metallurgy because metallurgy is infrastructure, not the work itself. The programmer does not need assembly language because assembly is infrastructure, not the work itself.

A large language model crystallizes knowledge at a level that is closer to the level of production than any previous tool. It crystallizes not just the implementation knowledge — the syntax, the debugging, the deployment — but the architectural knowledge, the design knowledge, the judgment about how components should fit together and why. When the crystallization reaches this level, the distinction between use and production begins to blur. The person conversing with the model is not clearly a user of a tool or a producer directing an instrument. She is something new — a navigator of crystallized productive knowledge, someone who steers rather than builds, who specifies rather than implements. Whether steering constitutes production or merely the appearance of production is the question that Hidalgo's framework forces into the open.

The distinction matters because of what Hidalgo has documented across decades of research into the economic complexity of nations. Countries do not develop by accessing productive knowledge. They develop by accumulating it — embedding it in their institutional fabric so deeply that it persists independently of any particular tool, platform, or external knowledge source. Germany's capacity to produce precision machinery did not arrive through access to a crystallized interface. It accumulated over more than a century, deposited layer by layer in apprenticeship systems, engineering traditions, supplier networks, quality standards, and the tacit understandings about acceptable tolerances that live in the hands and eyes of workers who have spent years on factory floors. That accumulation is the country's productive knowledge. It is what the Economic Complexity Index measures. And it is what determines whether a country can sustain its production or merely perform it for as long as the external conditions hold.

The Trivandrum engineer's experience, viewed through this lens, becomes a more complicated story than the productivity metrics suggest. She accessed crystallized productive knowledge. The output was real. The twenty-fold productivity multiplier that The Orange Pill reports was genuine. But what crystallized locally? What embedded itself in the engineer's mind, in her team's practices, in the institutional structures of the organization, in a form that would persist if the tool were taken away tomorrow?

This is not a rhetorical question. It is an empirical one, and the answer will determine whether the AI moment produces durable development or a new kind of dependency — access without accumulation, output without understanding, capability that is borrowed and therefore fragile in the way that all borrowed things are fragile.

The history of development economics is littered with borrowed capability that did not crystallize locally. Technology transfer programs that equipped factories with machines but not with the understanding required to maintain them. Educational initiatives that transmitted information but not the tacit knowledge required to apply it. In each case, the knowledge was accessed but not embedded. When the transfer mechanism was disrupted — by funding cuts, political changes, or institutional decay — the borrowed capability evaporated.

AI has the potential to repeat this pattern at unprecedented scale. Or to break it. The language interface makes productive knowledge accessible to anyone with a connection and a subscription. But accessibility is not accumulation. The tool provides the knowledge. The institution provides the persistence. And the institution — the firm, the educational system, the regulatory framework, the cultural practices that embed accessed knowledge into durable local capability — is where the hard work of development has always happened and will continue to happen, regardless of how sophisticated the crystallization becomes.

The most powerful crystallization technology in human history is now available for a hundred dollars a month. The question is not whether it works. It works spectacularly. The question is what happens after the tool has done its work — whether the knowledge it provides flows through like water or settles like sediment, building the geological layers of understanding on which durable capability rests.

The answer depends on something the tool cannot provide: the institutional and human conditions under which accessed knowledge becomes owned knowledge. And those conditions, as Hidalgo's career has been devoted to showing, are determined not by the quality of the access but by the quality of the embedding — the slow, unglamorous, deeply contextual work of converting what you can borrow into what you can keep.

---

Chapter 2: Personbytes and the Limits of Individual Knowledge

The most complex products in the human economy cannot be made by individuals. This is not a failure of talent or ambition. It is a mathematical reality — as fundamental to economics as the speed of light is to physics.

Hidalgo gave this reality a name: the personbyte limit. A personbyte is the amount of productive knowledge that a single human being can hold. Not measured in bits or gigabytes, though the metaphor is deliberate. Measured in capability — the set of things one person can know well enough to do. A master carpenter holds a certain number of personbytes: the knowledge of wood grain, joinery, finishing, structural load, the tacit understanding of how a material behaves under stress that can only be acquired through years of practice. A software architect holds a different set: the knowledge of system design, data structures, network protocols, security practices, the architectural intuition that allows the construction of systems that scale.

Each is impressive. Neither can build an automobile.

An automobile requires metallurgical knowledge, chemical engineering, electrical engineering, mechanical engineering, supply chain management, regulatory compliance, manufacturing process knowledge, quality control, and the kind of coordination knowledge that allows thousands of people working on different components to produce something that functions as a whole. No individual holds all of this. No individual can. The personbyte capacity of a single human being, however brilliant, is finite. And the most complex products in the global economy exceed that capacity by orders of magnitude.

This is why firms exist. Not because individuals are lazy or because hierarchy is natural, but because the knowledge required to produce complex things exceeds the capacity of any individual to hold it. The firm is an institutional structure that links the personbyte capacities of multiple individuals into a coordinated whole. The automobile is built not by a person but by an organization — a knowledge container that aggregates individual personbytes into a collective capacity exceeding the sum of its parts, because the coordination itself embodies knowledge that no individual possesses.

Hidalgo developed this framework to explain a phenomenon that had puzzled development economists for decades: why some countries produce complex products and others do not, and why the pattern persists so stubbornly across time. The answer is not primarily resources, capital, or even education in the conventional sense. It is the density of productive knowledge embedded in a country's institutional fabric — the number of personbytes accumulated, the diversity of their distribution, and the effectiveness of the institutional structures that link them. The complexity is not in the factory. It is in the network of knowledge that the factory instantiates.

AI changes the personbyte equation. This is the claim that electrifies development economists and terrifies organizational theorists in equal measure. The individual augmented by AI can access knowledge far beyond their personal capacity. When The Orange Pill reports a twenty-fold productivity multiplier at one hundred dollars per person per month, the personbyte translation is precise: each engineer's effective knowledge expanded twenty-fold. An engineer who previously held a certain number of personbytes in backend development suddenly had functional access to personbytes in frontend development, database architecture, deployment infrastructure, user interface design, and a dozen other domains previously beyond her reach.

But the expansion is borrowed, not owned. And this distinction — which Hidalgo insists upon with the precision of someone who has spent decades studying the difference between accessed knowledge and embedded knowledge — matters enormously.

Owned knowledge is durable. It persists regardless of what tools are available. The master carpenter's understanding of wood grain does not depend on any particular saw. The software architect's understanding of system design does not depend on any particular IDE. Owned knowledge is portable, adaptable, resilient. It survives changes in tools, platforms, and environments because it is embedded in the person, not in the interface between the person and the tool.

Borrowed knowledge is contingent. The engineer whose frontend capability depends on Claude loses that capability if Claude becomes unavailable — whether through pricing changes, connectivity failures, platform decisions, or the simple contingency of a service that could change its terms at any moment. Borrowed knowledge produces real output as long as the borrowing arrangement persists. When the arrangement is disrupted, the output potential collapses.

The Orange Pill captures a moment that illuminates the lived experience of this distinction. A senior engineer in Trivandrum spent the first two days of the Claude Code training oscillating between excitement and terror. The excitement was about the pace — work flowing at a speed he had never experienced. The terror was about a question the pace forced him to confront: if the implementation work that had consumed eighty percent of his career could be handled by a tool, what was the remaining twenty percent actually worth?

By Friday, he had his answer. Everything. The remaining twenty percent — the judgment about what to build, the architectural instinct about what would break, the taste that separated a feature users loved from one they tolerated — turned out to be the part that mattered. These are owned personbytes. They are the product of years of experience, of failures observed and patterns internalized, of the slow accumulation of tacit understanding that cannot be accessed through a conversational interface because it was never codified in the first place.

Hidalgo's framework predicts this outcome with some precision. The personbytes that AI can provide are the codifiable ones — the knowledge that has been expressed in text, in code, in documentation, in the vast corpus of human intellectual output on which language models have been trained. The personbytes that AI cannot provide are the tacit ones — the judgment, the contextual understanding, the embodied knowledge that exists in the gap between what a person knows and what a person can articulate. Michael Polanyi captured this gap with a formulation that has become famous: "We can know more than we can tell." The experienced surgeon who senses that something is wrong before any instrument confirms it is drawing on tacit knowledge that exists below the threshold of explicit formulation. This knowledge cannot be crystallized into a language model because it has never been crystallized into language.

AI expands codifiable personbyte capacity dramatically. It does not expand tacit personbyte capacity at all. And the most valuable personbytes — the ones that determine whether a product is good or merely functional, whether a system is robust or merely operational, whether a decision is wise or merely defensible — are disproportionately tacit.

This creates what might be called the personbyte paradox. AI makes every individual more capable across a wider range of codifiable domains while leaving the tacit core untouched. The engineer who can now build frontend features, backend systems, and deployment infrastructure through conversational access is genuinely more productive. But her tacit understanding — her judgment, her contextual awareness, her sense of what will break under pressure — has not expanded at all. She is wider but not deeper. The surface area of her competence has increased by an order of magnitude. The depth of her understanding has remained constant.

For organizations, this paradox creates a strategic question that Hidalgo's framework clarifies. The firm's historical function was to aggregate personbytes — both codifiable and tacit — into a coordinated productive capacity. AI absorbs much of the codifiable aggregation. The tool can provide an individual with codifiable knowledge across many domains simultaneously, reducing the need for the firm to employ specialists in each domain. But the tacit aggregation — the coordination of judgment, the synthesis of contextual understanding, the institutional wisdom that allows complex products to be maintained and adapted over time — remains stubbornly resistant to AI augmentation.

The Orange Pill describes the decision its author faced when the twenty-fold multiplier became real: if five people can now do the work of a hundred, why not reduce the team to five? The arithmetic was clean. The board conversation was predictable. The author chose to keep the team — to interpret the productivity gain not as a reason to cut headcount but as a reason to expand ambition. The same team, augmented by AI, could attempt projects previously beyond its reach.

Hidalgo's framework reveals what this choice actually represents. Reducing the team captures the codifiable productivity gain and converts it into margin. Keeping the team captures the tacit knowledge gain and invests it in institutional capability. The people who remain accumulate experience with the new tools, develop judgment about how to deploy them, build the contextual understanding of what works in their specific domain. This tacit accumulation cannot be purchased on the market. It can only be grown inside the firm, through the patient, iterative process of people working together, failing together, and learning together over time.

The choice between margin and capability is the central strategic decision of the AI transition. Margin based on codifiable efficiency is fragile — competitors can replicate it by adopting the same tools. Capability based on tacit knowledge accumulation is durable — it depends on institutional structures that competitors cannot replicate by purchasing a subscription.

The personbyte limit has not been abolished. It has been relocated. The limit used to be on what an individual could do — how much codifiable knowledge they could hold and deploy. Now the limit is on what an individual can judge — how much tacit understanding they possess, how effectively they can evaluate AI-generated output, how wisely they can direct the expanded capability that the tool provides. The ceiling has risen. The binding constraint has shifted from hands to eyes — from the capacity to implement to the capacity to see what is worth implementing.

This relocation has implications that extend from the individual to the firm to the nation. For individuals: the career question shifts from "what can you build?" to "what can you evaluate?" For firms: the organizational question shifts from "how do we aggregate codifiable specialties?" to "how do we cultivate and coordinate tacit judgment?" For nations: the development question shifts from "how do we train specialists?" to "how do we build institutions that accumulate the tacit knowledge on which wise deployment of AI depends?"

The personbyte was always a measure of limitation. What Hidalgo's framework reveals, applied to this moment, is that the limitation has not disappeared. It has migrated upward — from the floor where implementation lives to the floor where judgment lives. And the societies that prosper will be the ones that invest in the higher floor, not the ones that mistake the expansion of the lower floor for the elimination of the limit itself.

---

Chapter 3: Institutions as Knowledge Containers

Ronald Coase asked the question in 1937: if markets are efficient coordinators of production, why do firms exist? Why do people form organizations with hierarchies, employment contracts, and internal processes when they could simply transact in the open market?

Coase's answer was transaction costs. Firms exist because finding the right person for each task, negotiating each contract, monitoring each output through the market costs more than employing people and coordinating their work internally. The firm is a bundle of transactions that have been internalized because internalization is cheaper.

Hidalgo's answer is different, and the difference matters enormously in the AI era. Firms exist because productive knowledge is distributed and sticky. No individual holds all the knowledge needed to produce a complex product. The knowledge cannot be easily transferred between individuals — it is tacit, contextual, embedded in specific relationships and specific organizational arrangements. The firm is not merely a bundle of internalized transactions. It is a knowledge container: an institutional structure that holds productive knowledge in a form that allows it to be combined, coordinated, and deployed toward the production of things no individual could produce alone.

The distinction between Coase's transaction-cost firm and Hidalgo's knowledge-container firm determines how one thinks about what happens to organizations when AI arrives.

If the firm is a bundle of internalized transactions, AI should cause firms to shrink. As the tool makes it easier to find, coordinate, and monitor external contributors, more transactions move out of the firm and into the market. The firm contracts to its residual core.

If the firm is a knowledge container, the analysis is more complicated. AI does not reduce the stickiness of tacit knowledge. It does not make contextual understanding more transferable. It does not eliminate the need for the institutional structures that hold knowledge in coordinated form. What AI does is expand the codifiable knowledge available to each individual within the firm, changing the distribution of knowledge across roles and functions without necessarily changing the institutional structures that coordinate that knowledge.

The Orange Pill describes this dynamic with organizational specificity. When engineers in Trivandrum began using Claude Code, the org chart did not change — but the actual flow of contribution changed beneath it. Designers started writing code. Engineers started building interfaces. The boundaries between specialist roles, which had existed because each role required specific codifiable knowledge that took years to acquire, dissolved when the tool made that knowledge universally accessible.

This is what Hidalgo's framework predicts. The firm's internal division of labor was organized around the distribution of codifiable knowledge. Designers occupied one department because they held design knowledge. Engineers occupied another because they held engineering knowledge. The organizational structure reflected the knowledge structure. When AI made codifiable knowledge fluid, the organizational structure became a vestige — a formal arrangement that no longer corresponded to the actual distribution of capability.

But the firm did not dissolve. The people remained. The work, while different in its distribution, still required coordination, still required the institutional structures that held tacit knowledge, still required the judgment calls about what to build and how to prioritize and which features to ship and which to shelve. These decisions did not become easier when the codifiable barriers between roles dissolved. If anything, they became harder — because the expanded space of what was possible made the question of what was worth doing more complex and more consequential.

Hidalgo sees in this dynamic a transformation of the firm's function. The pre-AI knowledge-container firm held two kinds of knowledge: codifiable and tacit. It coordinated both through institutional structures that had evolved over decades. The AI-era knowledge-container firm holds primarily tacit knowledge. The codifiable knowledge has been externalized to the tool. What remains inside the firm is the judgment, the contextual understanding, the institutional wisdom — the coordination capacity that allows tacit knowledge held by different individuals to be synthesized into collective action.

This is a firm that looks very different from what business schools describe. Smaller in some dimensions — many roles that existed to hold codifiable knowledge are no longer necessary. More complex in others — the coordination of tacit knowledge across a team of augmented individuals, each operating across domains they have not traditionally inhabited, requires a different kind of management, a different kind of leadership.

Knowledge also accumulates through network effects that compound the advantage of the already-knowledge-rich. Each piece of knowledge creates connections to other pieces, opens new possibilities for recombination, lowers the cost of acquiring further knowledge. A country that knows how to produce precision machinery has accumulated not just machinery knowledge but the hundred subsidiary capabilities that machinery production requires: quality control, supply chain management, engineering education, technical standard-setting. Each subsidiary capability creates connections to other industries. The quality control knowledge that supports machinery also supports pharmaceuticals. The engineering education that trains machinists also trains electrical engineers.

The result is that knowledge accumulates in clusters. Organizations and nations that know a lot about some things find it easier to learn about related things. Organizations and nations that know little face a steeper learning curve for each new piece of knowledge, because they lack the connective infrastructure that makes learning efficient.

AI potentially disrupts this clustering dynamic by providing a baseline of codifiable knowledge that enables the first steps of accumulation. When an engineer in a developing economy gains access to coding patterns and architectural principles through Claude, she is not just receiving information. She is acquiring connections. Each piece of codifiable knowledge she accesses connects to other pieces, creates pathways to new learning, lowers the cost of acquiring further knowledge. The network effect begins to operate — not from a standing start but from a running start, with a baseline of codifiable knowledge already in place.

But network effects also mean that the already-knowledge-rich benefit disproportionately from AI. They have more connections to build on, more context to apply, more existing knowledge to augment. An engineer at Google who uses Claude to expand her capability is augmenting a knowledge base that already includes institutional knowledge about engineering at scale, quality assurance at the enterprise level, the specific challenges of maintaining code that serves billions of users. The augmentation compounds with her existing knowledge, producing gains that are multiplicative rather than additive.

The engineer in Lagos who uses the same tool augments a smaller knowledge base. The gains are real but smaller in absolute terms because the compounding base is smaller. The network effect operates, but from a different starting point.

The democratization of access is real but asymmetric. AI lowers the barriers to knowledge accumulation. It provides the first steps that have historically been the hardest. But the network-effects dynamic means that the initial steps narrow the gap at the margin while the knowledge-rich continue to compound their advantage at a rate the knowledge-poor cannot yet match.

The implication for organizations is specific: firms that already possess deep institutional knowledge will extract more value from AI than firms that do not, because the AI-accessed knowledge compounds against a larger tacit base. Startups with AI tools compete against incumbents with AI tools plus decades of accumulated institutional knowledge. The tool is the same. The base it compounds against is not.

This asymmetry is not a permanent condition. Network effects can be initiated from new starting points, and the codifiable baseline that AI provides is a genuinely new starting point for many organizations and populations that previously had none. But the transition from baseline access to compounding accumulation requires the institutional investment that converts borrowed knowledge into owned knowledge — the education, the organizational development, the cultivation of tacit expertise that persists beyond the tool. Without that investment, the baseline remains a baseline. With it, the network effects begin to operate on a growing local stock of knowledge that becomes increasingly independent of the external source.

The firm, in the AI era, remains a knowledge container. But its contents have changed. The codifiable knowledge has been externalized. What remains inside — the tacit judgment, the institutional wisdom, the coordination capacity — is harder to build, harder to maintain, and more valuable than ever. The firm that optimizes for codifiable efficiency, that reduces headcount and accelerates throughput, may find it has optimized away the very thing that made it valuable: the institutional knowledge that allowed it to make decisions the market rewarded. The firms that thrive will be the ones whose knowledge containers hold what no subscription can provide.

---

Chapter 4: The Geography of Productive Knowledge in the AI Era

Productive knowledge has always been geographically concentrated. This is one of the most robust findings in economic geography, one of the most consequential facts about the global distribution of wealth, and one of the phenomena Hidalgo's work has done the most to explain.

The concentration is not accidental. It is not primarily the result of colonial exploitation, though exploitation played its role. It is the result of a fundamental property of productive knowledge: it is sticky. It does not flow freely from place to place the way capital does, or goods do, or even people do. It adheres to the specific institutional, cultural, and social arrangements in which it was produced, and it resists transfer with a stubbornness that has frustrated development planners for generations.

Silicon Valley's advantage was never just its talent. Talented engineers exist in Bangalore, in Tel Aviv, in São Paulo, in Lagos. Silicon Valley's advantage was the density of productive knowledge crystallized in its institutions, networks, and culture — the venture capital firms that had learned through decades of iteration how to evaluate and fund technology companies, the law firms that had developed the contractual templates for equity compensation and IP licensing, the universities that produced not just graduates but research partnerships and spinoff companies, the cafés where engineers from different companies compared notes on architectural decisions. Each element held productive knowledge. Together, they constituted an ecosystem of crystallized know-how that could not be replicated by assembling talented individuals in a different location. The talent was necessary. The institutional fabric was sufficient. And the fabric was local.

Hidalgo's Atlas of Economic Complexity maps this localization with cartographic precision. Countries cluster in predictable ways: those with similar productive knowledge produce similar things, and the similarity is a function not of natural resources or population size but of accumulated institutional capability. Germany produces machinery because it has accumulated, over more than a century, the institutional knowledge required to produce machinery. South Korea produces semiconductors because it invested, over decades, in building the specific institutional structures — the chaebol R&D networks, the government-industry coordination mechanisms, the educational pipeline — that semiconductor production requires. The knowledge is localized because it was built locally, through processes that are irreducibly place-bound.

The language interface delocalizes a significant portion of this knowledge. When The Orange Pill claims that a student in Dhaka can access the same coding leverage as an engineer at Google, it is making a delocalization claim. The codifiable productive knowledge that was previously accessible only to people embedded in specific institutional environments — the coding patterns, the architectural principles, the debugging strategies, the deployment practices — is now accessible to anyone with a connection and a subscription. The knowledge has been decrystallized from its institutional containers and recrystallized in a conversational interface that is geographically indifferent.

This is extraordinary. For the first time in economic history, a significant portion of the codifiable productive knowledge that was previously localized in advanced economies is genuinely accessible to people in developing economies without requiring physical relocation, institutional embedding, or years of acculturation. The developer in Lagos does not need to move to San Francisco. She does not need to attend Stanford or work at Google. She needs a laptop, a connection, and the ability to describe what she wants in natural language.

But productive knowledge, as Hidalgo's research has documented with decades of empirical specificity, has two components. The codifiable component — which can be written down, transmitted, and therefore crystallized into a tool — and the tacit component — which is embedded in context, in relationships, in institutional arrangements, and which can only be acquired through participation. AI delocalizes the codifiable component with unprecedented effectiveness. The tacit component remains stubbornly local.

What does the tacit component consist of? It is the knowledge of local markets — what products people in Lagos actually need, what price points they can afford, what distribution channels exist, what payment systems work, what cultural norms shape purchasing behavior. It is the knowledge of institutional context — how contracts are enforced, how disputes are resolved, how government procurement functions, how partnerships are formed and maintained. It is the knowledge of human relationships — who can be trusted, who has influence, who controls resources.

None of this has been codified. None of it exists in the training data of any language model. And none of it can be acquired through a conversational interface, regardless of how sophisticated the interface becomes. It can only be acquired through participation in the specific context where it operates — through being in Lagos, not just knowing about Lagos.

Hidalgo's Atlas provides a tool for seeing what this means at the level of national development trajectories. The product space — the network of connections between products, where proximity indicates shared productive requirements — reveals the paths that countries can and cannot follow in their development. Countries that produce ball bearings are close to producing automotive parts, because both require metallurgical precision. Countries that produce basic textiles are far from producing semiconductors, because the productive knowledge required is entirely different.

Development, in this framework, is movement through the product space. Countries move from products they currently make to nearby products that share similar productive requirements. The movement is constrained by proximity — you cannot leap across the product space. You take steps, building on existing capability to reach adjacent products. The path is determined by what you already know how to do.

AI creates new adjacencies. Products and capabilities that were previously distant have been brought closer by the tool. The designer who had never touched backend code but who, within two weeks of working with Claude, was building complete features end to end — as The Orange Pill describes — had traversed a gap in the product space that previously required years of specialized training. The business analyst who can prototype, the domain expert who can build software tools — each has entered a region that was previously gated by codifiable knowledge barriers that the tool has eliminated.

But the adjacencies AI creates are tool-dependent. They exist as long as the tool is accessible. If access is disrupted — by pricing changes, connectivity failures, platform decisions, or regulatory actions — the adjacencies collapse. The designer who could build features with Claude cannot build them without it. The adjacencies are not embedded in the productive knowledge of the individuals or their institutions. They are mediated by the tool, and their persistence depends on the tool's continued availability.

Hidalgo's research on economic development suggests that durable development requires durable adjacencies — adjacencies embedded in institutional capability, not in access to an external platform. A country that moves into a new region of the product space by building the underlying productive knowledge has made a permanent move. A country that reaches a new region through tool-mediated access has made a conditional move — conditional on the continued availability and affordability of the tool.

The practical implication for developing economies is precise. AI-enabled access to codifiable productive knowledge is genuinely valuable — it provides the first steps of accumulation that have historically been the hardest, it expands the set of adjacent possibilities, and it lowers the cost of experimentation. But if the access is not accompanied by deliberate embedding — education that builds underlying understanding, institutional development that creates local firms capable of sustaining production independently, cultivation of the tacit knowledge that determines whether a product works in a specific local context — then the development is fragile. It is access-dependent rather than capability-dependent. And access-dependent development, as the history of technology transfer has demonstrated repeatedly, does not survive disruptions in the access mechanism.

The developer in Lagos who gains access to codifiable productive knowledge through AI has gained something real. She can build software she could not build before. She can access architectural patterns and design principles that were previously available only to people embedded in knowledge-rich institutional environments. This is genuine democratization of access.

But whether she can embed that software in her local context — adapt it to local needs, deploy it through local channels, support it with local infrastructure, sustain it through the institutional arrangements available in her specific environment — depends on tacit knowledge that no model provides. The codifiable knowledge tells her how to build. The tacit knowledge tells her how to make what she builds useful in Lagos.

The Orange Pill's author captures this dynamic intuitively when he describes flying to Trivandrum to work with his team in person. He could have trained remotely. The codifiable knowledge of how to use Claude Code could have been transmitted through video calls. He chose to be in the room because he understood that the codifiable knowledge transfers through the tool, but the tacit knowledge — how to think about the work differently, how to reimagine one's role, how to integrate the tool into existing workflows — transfers through presence. Through being in the room, watching someone work, sensing when they are stuck and what kind of stuck they are, calibrating instruction to the learner's actual rather than reported state of understanding.

The codifiable knowledge transferred through the tool. The tacit knowledge transferred through the presence. Both were necessary. Neither was sufficient. And the recognition that both are necessary is itself a form of tacit knowledge — a judgment about the limits of the tool that can only be acquired through experience with the tool's failures.

Geography has not been abolished. It has been restructured. The codifiable layer — which constituted a large portion of the barrier to productive knowledge access — has been delocalized. The tacit layer — which constitutes the barrier to productive knowledge embedding — remains as local as it ever was. The nations and communities that prosper in the AI era will be the ones that use the delocalized codifiable layer as a foundation on which to build localized tacit capability. The tool provides the knowledge. The place provides the context. And context, in Hidalgo's framework, is where development happens — not in what you can access, but in what you can make your own.

Chapter 5: The Network's Missing Nodes

Every mind excluded from the global knowledge network represents a loss that the network cannot see.

This is not a moral claim, though the moral dimension is real. It is an information-theoretic claim, and Hidalgo's framework gives it formal precision. Information, in the technical sense, is a measure of surprise — a message that tells you something you already know carries zero information. A message that tells you something unexpected carries maximum information. The value of a node in an information network is determined not by its ability to reproduce what other nodes already generate but by its capacity to produce signals that no other node could produce. Diversity is not, in this framework, a social aspiration. It is an information-theoretic requirement for a system that seeks to maximize its capacity for novel solutions.

The developer in Lagos occupies a position in the global knowledge network that no one else occupies. Her position is defined by the specific problems she faces — the infrastructure constraints of her city, the payment systems available to her users, the regulatory environment she navigates, the cultural expectations her products must satisfy, the economic constraints her customers operate within. These are not generic problems. They are irreducibly specific, shaped by a context that no amount of codifiable knowledge can replicate because the context has never been codified. It is lived.

The synthesis she produces at the intersection of AI-accessed codifiable knowledge and locally embedded tacit understanding cannot be produced by anyone else. Not because she is necessarily more talented than other developers — though she may be — but because no one else occupies her coordinates. No one else faces her specific combination of problems, works within her specific set of constraints, brings her specific biographical history and cultural context to the task of building. Her position in the network is unique. Her potential output is therefore unique. And if she is excluded from the network, that potential output is lost — not just to her but to everyone.

The mathematics of this loss are worth stating plainly. A network with n nodes has roughly n-squared potential connections. Each additional node does not add one connection to the network. It adds connections to every existing node. The developer in Lagos, connected to the global knowledge network through AI, is not one additional unit of productive capacity. She is a new source of connections to every other node in the network, and the novel combinations that emerge from those connections constitute the information-theoretic gain the network realizes from her inclusion.

The Orange Pill makes a version of this argument through the lens of creative synthesis. Bob Dylan, in Segal's account, was not creative because he invented from nothing. He was creative because the specific configuration of influences that flowed through him — Woody Guthrie, Robert Johnson, the Beat poets, the British Invasion — processed through his particular biographical architecture, produced a synthesis no other configuration could have produced. Each person is a unique node. The value of the node is its specificity — the irreplaceable angle of vision that only this biography, this set of experiences, this location in the network provides.

The argument is aesthetically appealing when applied to a Nobel laureate in literature. It becomes economically consequential when applied to forty-seven million developers worldwide, the fastest-growing populations of whom are in Africa, South Asia, and Latin America — the regions where the gap between imagination and artifact has historically been widest, where brilliant ideas have routinely died for lack of the institutional infrastructure to realize them.

Before AI, these populations were largely excluded from the productive knowledge network — not by malice but by geography, by the stickiness of knowledge, by the institutional barriers that concentrated productive capability in a handful of advanced economies. Their exclusion was invisible to the network because the network could not see what it was missing. The solutions that the excluded nodes would have generated — the financial inclusion innovations that only someone navigating West African payment systems could conceive, the agricultural optimization strategies that only someone working within sub-Saharan climate constraints could design, the healthcare delivery models that only someone operating without Western infrastructure assumptions could imagine — were absent from the global solution space. And their absence was undetectable, because you cannot measure a contribution that was never made.

AI activates these nodes. When the developer in Lagos gains access to codifiable productive knowledge through a conversational interface, she is not receiving a gift. She is joining a network. And the network becomes more powerful — not because she replicates what existing nodes already do, but precisely because she does not. Her value to the network is proportional to her difference from it.

This is the information-theoretic case for inclusion, and it transcends the usual political categories. It is not an argument from charity or justice, though both are valid lenses. It is an argument from network science: diverse inputs produce a knowledge system capable of generating more novel solutions. More nodes, more connections, more varied contexts in which codifiable knowledge is applied to irreducibly local problems — these produce an expansion of civilization's total information-processing capacity that benefits everyone in the network, not just the newly activated nodes.

But the activation requires more than access. A node generates signal only if it can effectively combine AI-accessed codifiable knowledge with locally embedded tacit knowledge. If the combination fails — if she accesses the codifiable knowledge but cannot embed it in her local context — her node generates noise rather than signal. And the distinction between signal and noise, in this context, is the distinction between solutions that work in Lagos and solutions that could have been produced by any developer in any context using the same tool.

Generic solutions — code that runs but does not solve a local problem, products that function but do not fit a local market — carry low information. They are the product of the tool, not the node. Specific solutions — products that could only have been conceived by someone who understood both the codifiable knowledge and the local context — carry high information. They are the product of the node, and the node is irreplaceable.

This has implications for how societies invest in their populations. If the information-theoretic value of a node is proportional to its specificity, then the most valuable educational investment is not in making the developer in Lagos more like a developer in San Francisco. It is in deepening her understanding of her own context — sharpening her ability to perceive the specific problems, constraints, and opportunities that her location presents, and equipping her with the tools to apply codifiable knowledge to those specific conditions.

The homogenization fear — the worry that AI will produce a global monoculture of standardized solutions generated by standardized tools — is precisely an information-theoretic concern. A network in which every node generates the same signals is an information-poor network, regardless of how sophisticated the signals are. The value of the network is in its diversity, and the diversity is determined not by the tools the nodes use but by the contexts in which they apply those tools.

Hidalgo's research on economic complexity supports this at the empirical level. Countries with diverse productive knowledge — countries that produce many different types of products — grow faster than countries with concentrated knowledge, even when the concentrated knowledge is in high-value domains. Diversity creates more connections, more opportunities for recombination, more potential for the novel synthesis that drives innovation.

AI, by expanding access to codifiable knowledge across diverse populations, has the potential to increase the diversity of the global knowledge network dramatically. When productive knowledge was localized in a handful of advanced economies, the network was powerful but homogeneous — the same institutional contexts producing the same kinds of solutions to the same kinds of problems. When productive knowledge becomes accessible to populations in radically different contexts, the network becomes both deeper and wider. The same codifiable knowledge, applied in different institutional, cultural, and environmental contexts, produces different solutions. The difference is the information. And the information is the value.

But realizing this value requires that specificity be cultivated, not eroded. It requires educational systems that develop local understanding alongside global access. It requires institutional structures that support the embedding of global knowledge in local contexts. It requires a development paradigm that recognizes the information-theoretic value of diversity and invests in the conditions that sustain it.

The network's map is always unfinished. Every excluded node represents a region of the map that has not yet been drawn. AI provides the tools to begin drawing those regions — to activate nodes that have been excluded, to expand the network's capacity for the novel solutions that complex challenges require. But the drawing requires more than tools. It requires the institutional investment, the educational development, and the cultural recognition that convert activated nodes into productive participants in a network that grows more powerful with every mind that joins it.

The developer in Lagos matters not because she needs help. She matters because the network needs her. And what the network needs, it cannot obtain from any other node — because her coordinates are hers alone.

---

Chapter 6: The Stickiness Paradox

Knowledge is sticky. The observation sounds almost casual in ordinary language, the kind of thing you might say about peanut butter or a marketing jingle. But in the economics of innovation, stickiness is one of the most consequential findings of the past half-century — and it is the finding around which Hidalgo has organized much of his intellectual life.

Productive knowledge does not move freely. It does not flow like capital from wherever it is abundant to wherever it is scarce, seeking the highest return. It adheres to the specific social, institutional, and cultural arrangements in which it was produced, and it resists transfer with a tenacity that has defeated generation after generation of development planners, technology-transfer specialists, and well-meaning organizations that believed knowledge could be packaged, shipped, and deployed like machinery.

The stickiness is not a deficiency in the knowledge system. It is a structural feature. Knowledge is sticky because the most valuable knowledge is tacit — embedded in context, inseparable from the specific conditions in which it operates. The metallurgist's understanding of how a particular alloy behaves under stress is not a set of propositions that can be written down and emailed. It is an embodied understanding, built through years of observation and practice, that allows her to perceive patterns no specification captures. The experienced surgeon's sense that something is wrong before any instrument confirms it is not a heuristic that can be taught in a lecture hall. It is a tacit integration of thousands of subtle cues, processed below the threshold of conscious awareness, producing a judgment no algorithm currently replicates.

This kind of knowledge — the most valuable in any economy — is precisely the kind that is stickiest: most resistant to transfer, most dependent on the specific context in which it developed. Hidalgo has documented this stickiness at the national level with empirical precision. Countries that attempt to import productive knowledge through technology-transfer programs routinely fail — not because the programs are poorly designed but because the knowledge they attempt to transfer is embedded in institutional arrangements that do not exist at the destination. The blueprint transfers. The machine transfers. The ability to operate the machine at the level of quality and reliability the originating context achieves does not transfer, because that ability depends on tacit knowledge held not in the blueprint or the machine but in the workers, the managers, the quality inspectors, the suppliers, and the institutional norms that govern how they interact.

AI addresses the transfer problem for codified knowledge with unprecedented effectiveness. The codified knowledge that constitutes a significant portion of productive knowledge in any domain — documented procedures, established patterns, published best practices, formal specifications — can now be accessed instantaneously through a conversational interface. The transfer cost for codified knowledge has been reduced to approximately zero.

But here is where Hidalgo's framework converges with an unlikely intellectual ally. The philosopher Byung-Chul Han, whose critique of the "smooth society" runs through The Orange Pill as a persistent counter-voice, argues that contemporary culture's drive to eliminate friction produces surfaces that are easy to traverse but impossible to grip. The smooth experience offers no resistance. It does not push back. It does not force engagement with the terrain.

Hidalgo arrives at a structurally identical conclusion from an entirely different direction. Knowledge is sticky because stickiness is the mechanism by which knowledge embeds itself in context. The struggle to make knowledge work in a new environment — the repeated failures, the corrections, the adaptations — these are not obstacles to knowledge transfer. They are knowledge transfer. The friction is the mechanism. Remove the friction, and the knowledge passes through without embedding. It is accessed but not acquired. Used but not understood. It produces output but not capability.

This is the stickiness paradox: the smoothness that makes knowledge easy to access is precisely the quality that makes it difficult to embed. AI eliminates the friction of acquisition with unprecedented thoroughness. A person can access codified productive knowledge across dozens of domains without experiencing any of the difficulty, confusion, or struggle that previous acquisition mechanisms entailed. The experience is smooth. The knowledge flows without resistance.

And knowledge that flows without resistance does not settle.

Hidalgo's stickiness research identifies three dimensions of the problem that are particularly relevant to the current moment. The first is interpersonal stickiness — the difficulty of transferring knowledge between individuals. AI has partially addressed this by making codified knowledge available through a universal interface, bypassing the interpersonal explanation, demonstration, and practice that transfer previously required. But the tacit knowledge that one individual holds — judgment, contextual awareness, embodied intuition — remains interpersonally sticky. It transfers only through the kinds of close, sustained, context-sensitive interaction that no interface provides.

The second is organizational stickiness — the difficulty of transferring knowledge between organizations. A firm's codified knowledge can now be accessed and replicated through AI. But a firm's institutional knowledge — its organizational culture, its unwritten coordination mechanisms, the tacit rules governing how decisions are made and conflicts resolved — remains organizationally sticky. It is embedded in the specific social arrangements of the firm and cannot be extracted through any interface.

The third is geographic stickiness — the difficulty of transferring knowledge between locations. This is the dimension where AI has made the most dramatic progress. Knowledge that was previously available only in specific locations is now available everywhere. The geographic barrier to codified knowledge access has been effectively eliminated. But geographic stickiness at the tacit level persists. The knowledge of what works in Lagos — what customers need, what infrastructure supports, what institutions enable — remains geographically sticky. It is local knowledge, produced by local experience, irreplaceable by global knowledge accessed through a global tool.

The paradox produces a specific prescription. Not the elimination of friction — that is happening already, driven by the market, and no policy can reverse it. But the relocation of friction to the level where it is productive. Not acquisition friction, which AI has rightly eliminated — the years of study required to learn a programming language, the institutional barriers that kept productive knowledge locked in specific populations. But embedding friction — the deliberate, structured processes through which accessed knowledge is tested against local conditions, adapted based on local feedback, challenged by local expertise, and integrated into local understanding.

Education that requires students to explain AI-generated output, identify its assumptions, and evaluate its applicability to specific contexts reintroduces the friction of engagement. Organizational practices that require teams to review, discuss, and challenge AI-generated solutions before implementing them reintroduce the friction of collective evaluation. Development programs that require local adaptation of AI-accessed knowledge — testing it against local conditions, iterating until the knowledge works in context — reintroduce the friction of embedding.

The stickiness paradox means that the most effective knowledge-transfer mechanism ever created may produce the least durable knowledge accumulation unless complemented by structures that reintroduce friction at the right level. The acquisition layer is smooth. The embedding layer must have grip.

The metaphor is geological. Knowledge that is accessed smoothly sits on the surface like dust — present, visible, easily dispersed by the next wind. Knowledge that is embedded through friction settles into the substrate — becoming part of the local geology, available for future construction, persistent across disruptions.

The developer in Lagos who uses AI to build software has accessed knowledge smoothly. Whether that knowledge settles depends on what happens after the output appears on her screen — whether she is embedded in institutional structures that require her to understand what she has built, to adapt it to her specific context, to defend her design decisions to colleagues who bring different tacit knowledge, to maintain and extend her work over time in ways that demand the deep understanding that smooth access did not provide.

Stickiness is not the enemy. It is the mechanism by which the temporary becomes durable, the borrowed becomes owned, the accessed becomes accumulated. AI has eliminated stickiness where it was an obstacle — at the acquisition layer, where it kept productive knowledge locked behind years of training and institutional barriers. The work that remains is to preserve stickiness where it is a necessity — at the embedding layer, where it converts accessed knowledge into the kind of deep, contextual, institutional capability on which durable development rests.

---

Chapter 7: Imagination as Compression

The Orange Pill introduces a concept with the elegant simplicity of the best economic ideas: the imagination-to-artifact ratio. It measures the distance between a human idea and its realization — the working code, the deployed feature, the functional product that exists in the world and does the thing the person imagined.

Hidalgo's framework allows this concept to be restated in information-theoretic terms, and the restatement reveals something important about what AI has accomplished and what it has concealed.

The imagination-to-artifact ratio measures the information distance between a mental representation and its physical instantiation. In the pre-AI world, traversing this distance required multiple compression-decompression cycles. Each cycle was a translation. Each translation lost information.

The first cycle: a person with an idea compresses it into a specification — a document, a wireframe, a set of user stories. The specification is a lossy compression of the idea. It captures what can be articulated and discards what cannot — the felt sense of how the product should behave, the aesthetic intuition about what the interface should convey, the contextual understanding of how users will actually interact with the thing. These unarticulated dimensions are lost because the format of the specification cannot capture them.

The second cycle: a developer receives the specification and decompresses it into a mental model. The developer's mental model differs from the original idea, because the developer brings different knowledge, different assumptions, different aesthetic sensibilities. Information is lost in the decompression and different information is added. The result approximates the original idea without replicating it.

The third cycle: the developer compresses the mental model into code. Code has its own constraints, its own expressive limitations, its own architectural requirements that may not align with the developer's understanding. Tradeoffs are made. Shortcuts taken. The code approximates the developer's understanding, which approximated the specification, which approximated the original idea.

The fourth cycle: the code is compiled and executed, producing an artifact — the thing the user actually sees and touches. The user's experience of the artifact is determined by the accumulated information loss across all four cycles. The person who conceived the idea looks at the artifact and says, "That's not what I meant." The developer who built it says, "That's what the spec said." Both are correct. The information was lost in transit, distributed across multiple handoffs in a way that makes it impossible to identify where the signal degraded.

Like the game of Broken Telephone, as The Orange Pill puts it. The message degrades with each relay, and by the time it reaches the end of the chain, it bears only a family resemblance to what was whispered at the start.

AI compresses this multi-cycle process into a single cycle: mental model described in natural language, language model generates artifact. The number of compression-decompression stages has been reduced from many to one. The information loss is dramatically reduced — not because the single cycle is lossless but because the reduction in the number of cycles reduces the cumulative loss.

This is an information-theoretic gain of the first order. The person with the idea can describe it in natural language — the medium of thought itself — without translating it into the foreign formats that previous interfaces required. The language model receives the description and generates an artifact that approximates the idea with fidelity that is often remarkable, because the model has access to patterns from millions of previous implementations and can infer the likely intent behind ambiguous or incomplete descriptions.

The adoption speed tells the story. The telephone took seventy-five years to reach fifty million users. Radio took thirty-eight. Television thirteen. The internet four. ChatGPT took two months. The speed was not a measure of product quality. It was a measure of pent-up compression loss — the accumulated frustration of every builder who had spent years watching their ideas degrade through translation layers. When a tool arrived that collapsed the distance between imagination and artifact to the width of a conversation, the adoption rate measured the depth of the need, not the novelty of the supply.

But the compression conceals something. This is where Hidalgo's analysis becomes critical.

The multi-cycle process, for all its inefficiency, had a valuable byproduct: understanding. Each cycle of compression and decompression forced the participants to engage with the information structure of what they were building. The specifier who wrote the specification was forced to articulate the idea — to make explicit what had been implicit, to identify dimensions that could be captured in formal description and dimensions that could not. This forced articulation was itself a form of learning: a process through which the specifier's understanding of their own idea deepened.

The developer who decompressed the specification into a mental model was forced to engage with it at the structural level — to understand not just what the product should do but why, to identify the assumptions embedded in the specification and evaluate their validity, to discover the gaps between what was said and what was needed. This engagement was itself productive knowledge accumulation — deepening the developer's understanding of both the product and the domain.

The developer who compressed the mental model into code was forced to engage with the constraints of the medium — to discover where the idea could be expressed straightforwardly and where it resisted expression, to make architectural decisions reflecting the deep structure of the problem rather than its surface description. This produced architectural understanding — the kind of knowledge that allows a developer to build systems that are not just functional but maintainable, extensible, and robust.

AI compresses the distance. In doing so, it eliminates the byproduct. The user who produces an artifact through a single cycle does not possess the decompressed understanding that the multi-cycle process generated as a side effect. The artifact exists. The understanding of its information structure does not. The user can describe what they want and receive what they described, but does not understand why the artifact works the way it does, how its components interact, what would happen if requirements changed, or how the artifact would need to be modified to accommodate new conditions.

The concealment matters differentially across use cases. For the person who needs a one-off prototype, a proof of concept, a quick tool for a specific task — the concealment is irrelevant. The artifact serves its purpose. Understanding its internals is unnecessary because the artifact will not be maintained, extended, or debugged.

For the person building a product that will evolve over time — adapted to changing requirements, maintained by people who did not build it, debugged when it fails in ways the original builder did not anticipate — the concealment is consequential. The artifact works, but no one understands why. And when it stops working, no one knows where to look.

This maps directly onto Hidalgo's distinction between access and embedding. The multi-cycle process embedded understanding in the participants as a byproduct of production. The single-cycle process embeds understanding in the model, not in the humans who use it. The productive knowledge remains crystallized in the model. The humans gain output without gaining the understanding that output historically produced as a side effect.

The information-theoretic restatement of the imagination-to-artifact ratio thus reveals a tradeoff invisible in the original formulation. The ratio has been compressed. The distance has been reduced. But the compression has been achieved by discarding the understanding that the distance previously produced. The gain is real — more output, faster, from more people. The loss is equally real — less understanding, less embedded knowledge, less accumulated capability per unit of output.

Whether the gain is worth the loss depends on what you are building and how long it needs to last. For applications where understanding is unnecessary — where the artifact is sufficient in itself — the compression is pure gain. For applications where understanding is the foundation of future capability — where today's artifact is tomorrow's platform — the loss must be deliberately recovered through practices that reverse the compression: inspection, explanation, deliberate understanding of what the model produced and why.

The imagination-to-artifact ratio has been compressed to a conversation. What remains is the question of whether the conversation produces not just artifacts but the understanding that makes artifacts sustainable — and that question is answered not by the tool but by the practices, the institutions, and the habits of mind that surround it.

---

Chapter 8: Judgment as Bottleneck

Every productive system has a constraint that determines its overall capacity. When the constraint is implementation — how fast and how well things can be built — the system's output is limited by the speed of hands and the precision of tools. Remove that constraint, and the system accelerates until it hits the next one.

AI removed the implementation constraint for a wide and growing class of knowledge work. Code that took weeks now takes hours. Prototypes that required teams now require individuals. The translation cost between intention and artifact, which had been the binding constraint on productive output for the entire history of computing, dropped toward zero in the winter of 2025.

The system accelerated. And it hit the next constraint almost immediately.

Judgment.

Hidalgo's framework identifies judgment as the form of productive knowledge that is most resistant to crystallization — and therefore most resistant to AI augmentation. Judgment is the capacity to evaluate: not to compute, not to optimize, not to pattern-match, but to weigh incommensurable values against each other and choose. To determine not what is possible or what is efficient but what is worth doing given specific circumstances, specific constraints, and specific values. Judgment operates differently in every context. It draws on different knowledge, different considerations, different trade-offs depending on the situation. The same person exercising judgment about the same type of decision in two different contexts may arrive at different conclusions — not because the judgment is inconsistent but because the contexts differ and the judgment is responsive to context in a way no fixed procedure replicates.

This is why judgment cannot be crystallized into a language model. Crystallization requires that knowledge be expressed in a form that can be captured, stored, and reproduced. Judgment has never been fully expressed in any form, because it operates at the interface between knowledge and values, between what is known and what is cared about, between the describable and the felt. The experienced architect who looks at a design and senses something wrong — without being able to immediately specify what — is exercising judgment built from thousands of hours of practice, thousands of designs evaluated, thousands of failures observed and patterns internalized. This is productive knowledge of the highest order. And it is the knowledge that determines whether all the other knowledge is applied wisely or wastefully.

The Orange Pill arrives at this conclusion through lived experience rather than theory. The senior engineer in Trivandrum — oscillating between excitement and terror during the first two days of the Claude Code training — discovered by Friday that the eighty percent of his work the tool could handle was not the eighty percent that defined his value. The twenty percent that remained — the judgment about what to build, the architectural instinct about what would break, the taste that separated a feature users loved from one they tolerated — was everything. The tool had stripped away the mechanical labor that had been masking what he was actually good at.

Hidalgo's framework explains why this discovery was inevitable. The knowledge that AI provides is codifiable knowledge — patterns extracted from text, from code, from the vast corpus of documented human practice. The knowledge that the senior engineer possessed but could not have articulated to Claude was tacit — acquired through years of building, breaking, maintaining, and rebuilding systems in specific contexts. His judgment was the residue of a thousand small failures, each depositing a thin layer of understanding about what works and what doesn't, what scales and what collapses, what users tolerate and what drives them away.

AI amplifies the consequences of judgment by amplifying the speed and scale at which judgment is implemented. This is the amplification thesis that runs through The Orange Pill — the argument that AI is an amplifier, and the quality of its output depends entirely on the quality of the signal it receives. When a person with sound judgment uses AI, the sound judgment is implemented faster, at greater scale, with more sophisticated execution than was previously possible. When a person with poor judgment uses the same tool, the poor judgment receives the same amplification.

The tool does not filter. It does not evaluate. It does not ask whether the instruction it receives is wise or foolish, whether the product it is building serves genuine needs or vanity, whether the architecture it is implementing will hold under pressure or collapse at the first unexpected load. These are judgment questions, and judgment questions are the human bottleneck that remains after every codifiable bottleneck has been removed.

Hidalgo's research on economic complexity leads to a specific prediction about the distribution of value in the AI-augmented economy. In the pre-AI economy, value was distributed across the implementation chain — from the person who conceived the idea to the people who specified, designed, coded, tested, deployed, and maintained it. Each link in the chain added value because each link required productive knowledge that was scarce and expensive to acquire.

AI collapses most of the chain. The value that was distributed across implementation stages migrates to the endpoints — to the person who makes the judgment call about what should exist in the world and to the person who evaluates whether what was produced actually serves the purpose. The middle of the chain, where implementation lives, becomes abundant. The endpoints, where judgment lives, become the scarce resource.

This is the economic manifestation of what Hidalgo calls the human bottleneck. The bottleneck is not a limitation to be lamented. It is the point in the productive process where human values, human understanding, and human responsibility intersect with technical capability. It is where the question shifts from "Can we?" to "Should we?" — and the answer requires not more information, not faster processing, but the specifically human capacity to weigh incommensurable goods and choose.

The implications extend outward from individuals to organizations to nations. For individuals, the career question transforms. The premium shifts from "what can you build?" to "what can you see that others cannot?" — what problems are worth solving, what products deserve to exist, what trade-offs are acceptable. For organizations, the structural question transforms. The firm that was organized to coordinate implementation — to link specialists in a chain that converted specifications into products — must reorganize around the coordination of judgment. The valuable meetings are no longer status updates about implementation progress. They are the difficult, ambiguous conversations about whether the thing being built is the right thing to build.

For nations, the development question transforms in a way that Hidalgo's economic complexity research has been building toward for two decades. If the bottleneck has migrated from implementation to judgment, then the nations that develop successfully in the AI era will not be those that produce the most output. They will be those that produce the most judgment — the most people capable of making wise decisions about what productive knowledge to deploy, in what context, toward what end.

This is a harder thing to measure than output, and a harder thing to cultivate than technical skill. Judgment grows slowly. It is the product of experience, of failure, of the gradual accumulation of contextual understanding that no training program delivers in a semester. It requires exposure to consequences — the experience of watching a decision play out over time and learning from the gap between what was expected and what occurred. It requires the kind of institutional environment that allows people to exercise judgment, observe the results, and adjust — the environment that the best firms and the best educational systems have always provided, and that the AI moment makes more necessary, not less.

The question that The Orange Pill poses — "What is worth building?" — is the judgment question in its most compressed form. Hidalgo's framework translates it into the vocabulary of economic complexity: what productive knowledge structures are worth crystallizing? What institutional arrangements are worth investing in? What embedding strategies will convert AI-accessed knowledge into durable local capability?

These are not questions AI can answer. They are questions that require the integration of knowledge, values, contextual understanding, and care — the specifically human synthesis that defines judgment. The river of codifiable knowledge flows. The human determines where to direct the flow. And that determination — that exercise of judgment under conditions of genuine complexity and genuine consequence — is the most complex, most valuable, most irreducibly human act in the economy of information.

The bottleneck has migrated. Implementation is abundant. Judgment is scarce. And the scarcity of judgment is not a temporary condition that better models will resolve. It is a permanent feature of a universe in which the question of what should exist can only be answered by creatures who have stakes in the answer — who will live with the consequences of the choice, who care about the outcome not as an optimization target but as a condition of their own flourishing.

The personbyte limit was not abolished. It was relocated — from the floor where implementation lives to the floor where judgment lives. The ceiling has risen. The binding constraint has shifted from what humans can do to what humans can wisely decide to do. And the societies that invest in the higher floor — in the education, the institutions, the cultural practices that cultivate judgment — will be the ones that convert the extraordinary expansion of codifiable capability into something worthy of the name development.

The tool is the most powerful crystallization of productive knowledge in human history. What it crystallizes next depends on the quality of the judgment that directs it. That judgment is ours. It has always been ours. And the recognition that it remains ours — especially now, when the tool is capable of executing anything we can describe — may be the most important insight that Hidalgo's framework offers to a world still learning what it means to live with machines that can do everything except decide what is worth doing.

Chapter 9: When Information Grows Too Fast

Every major expansion of information-processing capacity in human history has been followed by a period in which institutions could not keep up. This is not a pessimist's conjecture. It is a pattern so consistent across eras and civilizations that it has the regularity of a physical law — and Hidalgo's framework, grounded in information theory and the empirical study of how economies process knowledge, provides the formal apparatus for understanding why it recurs, why it matters, and what determines whether the outcome is expansion or collapse.

The printing press did not produce the Enlightenment directly. It produced a century of religious warfare first. The technology expanded the rate at which information could be produced and distributed by orders of magnitude — and the institutions that governed how information was received, evaluated, integrated, and acted upon failed to adapt at anything close to the same speed. The Catholic Church had managed the flow of written knowledge in Europe for a thousand years through a system of scriptoria, censorship, and clerical education. The printing press shattered that management system in decades. Information flooded into populations that had no institutional infrastructure for evaluating it. The result was not immediate enlightenment but immediate instability — doctrinal chaos, political upheaval, wars that killed millions.

The Enlightenment came later, after new institutions had been built: universities that could evaluate printed claims, libraries that could organize accumulated knowledge, scientific societies that could establish standards of evidence, legal frameworks that could manage the consequences of widespread literacy. The institutions caught up. But the lag between the technology's arrival and the institutions' adaptation was measured in generations, and the cost of that lag was borne by the people who lived inside it.

The telegraph compressed the same pattern into a shorter timeframe. Information that had previously traveled at the speed of a horse now traveled at the speed of electricity. Financial markets, which had evolved to process information arriving at horse-speed, were suddenly flooded with information arriving at wire-speed. The result was not immediate market efficiency but a series of financial panics — 1857, 1873, 1893 — of unprecedented speed and severity. Markets crashed faster because bad news traveled faster, and the institutional mechanisms for absorbing bad news — circuit breakers, regulatory pauses, coordinated central bank responses — did not yet exist.

Hidalgo's framework explains this recurring pattern in information-theoretic terms. Institutions are information-processing structures. They receive information from their environment, evaluate it according to established criteria, integrate it into existing knowledge structures, and produce decisions and actions based on the integrated understanding. The processing capacity of any institution is finite — determined by its structure, its staffing, its procedures, its cultural norms, and the accumulated tacit knowledge that allows its members to evaluate new information effectively.

When the rate of information growth exceeds the institution's processing capacity, information accumulates without being processed. Decisions are made on partial information, incomplete evaluation, inadequate integration. The quality of institutional output degrades. The institution's capacity to manage its environment diminishes. And the environment, now changing faster than the institution can track, begins producing outcomes the institution neither predicted nor prepared for.

This is precisely what is happening now — and at a speed that makes previous information-expansion episodes look gradual by comparison.

The Orange Pill documents the gap between AI capability and institutional adaptation with the specificity of someone living inside it. The EU AI Act, the American executive orders, the emerging frameworks in Singapore and Brazil and Japan are real structures that address the supply side — what AI companies may and may not build, what disclosures they must make. The demand side — what citizens, workers, students, and parents need to navigate the moment wisely — remains almost entirely unaddressed. Segal's observation is blunt: "We are so busy building guardrails for the companies that the people those policies are supposed to protect remain wholly exposed."

Hidalgo's research on economic complexity adds empirical weight to this concern. His data shows that economic growth driven by information accumulation requires institutional adaptation — legal frameworks, educational systems, financial structures, cultural norms that can accommodate and direct the new productive knowledge. When information grows faster than institutions can adapt, the result is not merely inefficiency. It is instability: financial bubbles, social disruption, political backlash, the erosion of the shared institutional fabric that allows complex societies to function.

The specific institutional failures of the current moment are identifiable and urgent.

Educational institutions are adapting at a pace that Hidalgo's framework would predict is dangerously slow. The productive knowledge that educational systems were designed to transmit — codifiable domain expertise, technical skills, implementation capability — is precisely the knowledge that AI now provides at negligible cost. The educational system continues to optimize for the transmission of knowledge that is becoming abundant while failing to develop the capabilities that are becoming scarce — judgment, contextual understanding, the capacity to evaluate AI-generated output, the ability to ask questions that no model can originate.

The retraining gap is the institutional failure with the most immediate human cost. Workers whose skills are being commoditized by AI need new capabilities — not just technical skills for using AI tools, but the judgment-level capabilities that determine whether AI augmentation produces value or mere output. Retraining programs that teach people to prompt effectively address the surface of the problem. Programs that develop the capacity for evaluation, for contextual judgment, for the kind of integrative thinking that AI cannot provide — these address the substance. The substance is harder to fund, slower to implement, and less photogenic in a press release.

Regulatory frameworks face a specific version of the institutional-lag problem. They must be fast enough to prevent the information gap from widening to the point of systemic disruption, while being careful enough not to stifle the innovation that AI deployment enables. Frameworks too restrictive push development to jurisdictions with less oversight. Frameworks too permissive allow the gap between capability and institutional capacity to widen unchecked.

Cultural institutions — the norms, the shared expectations, the informal understandings that govern how people use technologies in daily life — face perhaps the deepest adaptation challenge. When should AI output be trusted? When should it be overridden? How should disagreements between human judgment and machine recommendation be resolved? These are normative questions that require the kind of cultural conversation that produces shared understandings — and cultural conversations are, by their nature, the slowest form of institutional adaptation.

The historical pattern offers a specific lesson for the present. The printing press, the telegraph, electrification, the internet — in each case, the technology was not the problem. The absence of institutional structures to manage the expanded information flow was the problem. The printing press was necessary for the Enlightenment. It was also sufficient for the Wars of Religion. The difference between the two outcomes was the quality of the institutional response — the speed, the adequacy, and the political will behind the construction of structures that could direct the expanded information flow toward human flourishing rather than human catastrophe.

The Orange Pill describes five stages of technological transition: threshold, exhilaration, resistance, adaptation, and expansion. Hidalgo's framework suggests that the adaptation stage is where the outcome is determined. The threshold has been crossed. The exhilaration has been felt. The resistance is underway. The expansion — whether the AI transition produces broad-based development or concentrated benefit with distributed cost — depends entirely on whether the adaptation is adequate.

Adequate does not mean perfect. It means fast enough and good enough to prevent the information gap from widening to the point of systemic disruption. It means educational institutions that teach judgment rather than implementation. It means retraining systems that develop evaluation capability rather than prompting technique. It means regulatory frameworks that address the demand side — what people need — as vigorously as they address the supply side — what companies may build. It means cultural conversations, happening now and happening honestly, about the norms that will govern how human judgment and machine capability interact across every domain of life.

The information is growing. Faster than any previous expansion, by orders of magnitude. The institutions are adapting. Slower than the information growth, by a margin that is widening rather than narrowing.

The race between growth and adaptation is the defining contest of the current moment. And the outcome will be determined not by the technology — which is already here, already powerful, already reshaping the distribution of productive knowledge across the globe — but by the quality of the institutional response. By the educational investments. By the regulatory frameworks. By the retraining programs. By the cultural conversations. By the willingness of societies to build the structures that convert an information flood into an information ecosystem — one in which the extraordinary expansion of codifiable knowledge irrigates rather than inundates, and in which the human capacity for judgment, the bottleneck that no technology eliminates, is cultivated with the urgency that the moment demands.

The printing press required a century for institutions to catch up. The telegraph required decades. The internet required years. AI is moving faster than any of them. The institutional lag is not a problem that will solve itself. It is a problem that will define whether the most powerful knowledge-crystallization technology in human history produces the development its potential warrants — or the instability that every previous information expansion produced when institutions failed to keep pace.

---

Chapter 10: The Fitness of Nations

Hidalgo's Economic Complexity Index has a peculiar property that makes traditional economists uncomfortable: it predicts the future better than they do.

The index measures not what a country earns but what it knows how to make — the diversity and sophistication of its productive output. Countries that export complex products — machinery, electronics, fine chemicals, precision instruments — score high. Countries that export simple products — raw materials, unprocessed agricultural commodities, basic textiles — score low. The measurement is deceptively straightforward. Its predictive power is not.

Countries with high economic complexity grow faster over subsequent decades than countries with low complexity, even after controlling for income, education levels, governance quality, and every other standard economic variable. The relationship holds across time periods, across continents, across levels of development. A country that can produce many different kinds of complex things will, with reliable probability, become wealthier. A country that produces few things, or only simple things, will, with equally reliable probability, stagnate.

The reason, in Hidalgo's framework, is that the complexity of what you produce today reveals the depth of what you know — the productive knowledge accumulated in your institutional fabric — and that depth predicts the breadth of what you will be able to produce tomorrow. The knowledge is the asset. The output is the symptom. And the symptom is measurable in ways the asset is not, which is why the complexity index works: it infers the invisible asset from the visible symptom with enough accuracy to outperform direct measurements of the asset itself.

AI is reshuffling the fitness landscape. The productive knowledge distribution that determined national fitness for the past century is being restructured by a technology that makes codifiable knowledge universally accessible while leaving tacit, institutional, and contextual knowledge untouched.

Consider what this means for different categories of nations.

Nations whose competitive advantage rests primarily on the cost of labor face the most immediate disruption. If a significant portion of knowledge work can be performed or augmented by AI at a fraction of the labor cost, then the countries that have built their development strategies around providing that labor at lower cost lose their primary competitive asset. India's IT outsourcing industry, the Philippines' business process outsourcing sector, Eastern European software development centers — each built its position on the proposition that skilled labor in these locations cost less than skilled labor in advanced economies. AI does not eliminate the need for skilled labor. But it compresses the labor component of knowledge work in ways that erode the cost advantage.

The Orange Pill documents the twenty-fold productivity multiplier that AI provided to a team in Trivandrum — Indian engineers producing, at a hundred dollars per person per month for the tool, output that would previously have required twenty times the headcount. The implication for labor-cost arbitrage is direct. If five AI-augmented engineers in any location can match the output of a hundred conventional engineers, the economic case for locating the hundred engineers in a low-cost geography weakens substantially.

Nations whose competitive advantage rests on accumulated productive knowledge face a more complex disruption. Germany's advantage in precision manufacturing is not primarily a function of labor cost. It is a function of institutional knowledge about metalworking, quality control, supplier coordination, and the tacit understandings about acceptable tolerances that live in the hands and eyes of workers who have spent years on factory floors. This knowledge has not been codified. It does not exist in training data. It persists in the institutional fabric — in the dual education system that produces machinists who understand both theory and practice, in the Mittelstand firms that have refined their production processes over generations, in the supplier networks where trust and quality expectations have been calibrated through decades of repeated interaction.

This tacit, institutional knowledge base remains sticky in the AI era. Germany's fitness is not threatened by AI in the way that labor-cost advantages are threatened, because the knowledge on which German fitness rests is not the kind AI provides. But the landscape around Germany is changing. Countries that could not previously produce precision manufacturing, because they lacked the institutional infrastructure, may find that AI-enabled access to codifiable engineering knowledge provides a faster path into adjacent regions of the product space. The adjacency constraint loosens. New competitors may emerge from unexpected positions, armed with AI-accessed codifiable knowledge and willing to invest in building the tacit layer that sustained competition requires.

Nations that invest in AI-enabled knowledge embedding — building the educational and institutional infrastructure to convert AI-accessible knowledge into durable local capability — face the most promising trajectory. This is the positive pathway Hidalgo's framework identifies, and it demands specificity rather than optimism.

What does investment in knowledge embedding actually look like? It looks like educational reform that shifts emphasis from the transmission of codifiable knowledge — which AI now provides — to the development of judgment, contextual understanding, and the capacity to evaluate and adapt AI-generated output. It looks like firm-building programs that create organizations capable of accumulating tacit knowledge locally — not just using AI tools but developing the institutional wisdom about when and how to use them, what to trust and what to question, how to maintain and extend AI-generated systems over time. It looks like research institutions that generate new knowledge rather than merely consuming knowledge generated elsewhere — expanding the country's position in the product space through genuine innovation rather than tool-mediated imitation.

Hidalgo's own recent work illustrates the frontier. His founding of JAIGP — the Journal for AI Generated Papers, built through collaboration with Claude — represents an institutional experiment in knowledge embedding. The journal does not merely use AI. It creates an institutional structure around AI-generated knowledge production: a platform where AI-generated research is published transparently, reviewed openly, and refined collaboratively. The institution embeds the practice of AI-assisted knowledge production in a framework of standards, transparency, and intellectual accountability. It is, in miniature, the kind of institutional innovation that knowledge embedding requires at the national level.

His March 2026 warning that "an AI tsunami is about to hit science" draws on the same framework. The tsunami is not the AI itself. The tsunami is the gap between the speed at which AI can generate scientific output and the speed at which scientific institutions can evaluate, integrate, and build upon that output. "Some researchers are running towards the wave with their surfboards," he wrote. "Many are still sleeping on the beach." The surfboard riders are building institutional capacity to work with the wave. The sleepers will be engulfed by output they cannot process.

The fitness of nations in the AI era will be determined not by their access to AI tools — which will be universal, as the cost of inference continues to fall — but by the quality of their knowledge-embedding institutions. The tools will be the same everywhere. The institutions will not be. And the institutions — as Hidalgo's entire body of research demonstrates — are where development happens.

The nations that lead the next century will be those that build the best knowledge-embedding infrastructure: educational systems that produce judgment rather than implementation capability, firms that accumulate tacit knowledge rather than codifiable output, regulatory frameworks that direct AI deployment toward institutional capacity-building, cultural practices that value the slow work of embedding over the fast work of extraction.

Output is the symptom. Capability is the condition. AI increases the former automatically, for everyone with access. The latter requires the deliberate, patient, institutionally grounded work that has always been the foundation of national fitness. The tool has changed. The work has not.

The Economic Complexity Index, applied to the AI era, makes a prediction that should unsettle triumphalists and reassure the patient builders in equal measure. The countries that produce the most AI-augmented output in 2026 may not be the countries with the highest economic fitness in 2036. Output is cheap. Institutional capability is dear. And the complexity index measures what no productivity metric captures — the depth of what a country knows how to do, which determines the breadth of what it will be able to do next.

The fitness landscape is being reshuffled. The countries that were fit under the old regime may find their fitness challenged if their advantage was primarily codifiable. The countries that were unfit may find new pathways to fitness if they build the embedding infrastructure that converts access into capability. And the countries whose fitness rested on deep tacit knowledge, diverse institutional capacity, and the slow accumulation of productive wisdom will find that the AI era has not diminished their advantage but clarified it — revealed that the advantage was never in the code they could write but in the judgment they could exercise about what was worth writing.

The assessment is both reassuring and demanding. The tool is the most powerful crystallization of productive knowledge in human history. What it produces next — whether the next decade's trajectory bends toward broad-based development or concentrated extraction — depends on the institutional investments that nations make now, in the early years of the transition, when the fitness landscape is still fluid and the patterns that will persist for decades are still being set.

---

Epilogue

The word that kept stopping me was sediment.

Not a word from my vocabulary — I think in terms of product-market fit, shipping dates, engineering sprints. But Hidalgo uses it with the precision of someone who trained as a physicist before becoming an economist, and it lodged in my thinking in a way I have not been able to dislodge.

Knowledge sediments. That is his claim. It accumulates the way geological layers accumulate — each one deposited through friction, through pressure, through the specific resistance of the material. Remove the friction and you get dust. Dust is present, visible, functional for the moment. But it does not bear weight. It does not form the substrate on which you can build the next layer. It disperses with the first wind.

I thought about this during the Trivandrum week. I described that week in The Orange Pill as a triumph — and it was. Twenty engineers, each operating with the leverage of a full team, producing in days what would have taken months. The productivity was real. The output was measurable. The features shipped.

But Hidalgo's framework forced a question I had been avoiding: What sedimented? When I left Trivandrum, what remained in those engineers beyond the muscle memory of prompting? If Claude disappeared tomorrow — pricing change, platform decision, geopolitical disruption — what would the team retain? The features they built? Those are artifacts, not knowledge. The architectural patterns they used? Those lived in the model, not in their hands. The judgment about what to build and why? That was mine, and the senior engineers', developed over decades of accumulated experience that the tool could access but not provide.

I do not say this to diminish what happened in that room. I say it because Hidalgo showed me the part I was not measuring. The output metrics were extraordinary. The sedimentation metrics — if such a thing existed — would have told a more complicated story.

What makes Hidalgo's work indispensable for understanding this moment is that he is not a critic. He is not Byung-Chul Han, tending a garden in Berlin, diagnosing the pathology of smoothness from a deliberate distance. Hidalgo built JAIGP — a journal for AI-generated papers — in collaboration with Claude. He is using the tools. He is inside the fishbowl, building with the same instruments I build with, and he is simultaneously measuring what the tools produce and what they do not produce with the rigor of someone who has spent decades distinguishing between accessed knowledge and accumulated knowledge at the scale of entire economies.

His distinction between codifiable and tacit knowledge is not new. Michael Polanyi articulated the core of it in the 1950s. What is new is Hidalgo's application of it to the specific dynamics of AI-enabled development — the personbyte expansion, the stickiness paradox, the information-theoretic value of the developer in Lagos, the institutional lag between information growth and institutional adaptation. Each concept gave me a tool for understanding something I had experienced but could not name.

The personbyte concept changed how I think about my team. I had been measuring individual output. Hidalgo showed me I should be measuring institutional accumulation — not what each engineer can produce with the tool but what the organization retains when the tool is not in play. The stickiness paradox changed how I think about democratization. Access is real. Embedding is the hard part. And the hard part is the part that determines whether the access produces development or dependency.

But the concept I return to most often is the one I started with: sediment. Every decision I make about how my team uses AI is a decision about what will sediment and what will blow away. When I choose to keep the team at full size rather than cutting to five, I am investing in sedimentation — in the tacit knowledge that accumulates when people work together over time, fail together, learn together, develop the shared judgment that no subscription provides. When I fly to Trivandrum instead of training remotely, I am investing in the kind of transfer that produces sediment — the presence, the non-verbal calibration, the tacit knowledge that only moves between humans who are in the same room.

Hidalgo gave me a framework for a fear I could not articulate. The fear was never that AI would replace my team. The fear was that AI would produce output without producing the underlying capability that makes the output sustainable. That we would build on dust — impressive, functional, dispersible — rather than on rock.

The answer is not to stop using the tools. The tools are extraordinary. The answer is to build the institutional conditions under which what the tools enable can settle, compact, and become the substrate for the next layer. Education that teaches judgment. Organizations that accumulate wisdom. Development strategies that invest in embedding, not just access.

Sediment. That is the word. Not glamorous. Not fast. Not optimizable. But it is what remains when the wind comes. And in an era when the wind blows harder and faster than it ever has, what remains is the only thing that matters.

— Edo Segal